Yuvraj Raulji | AI Search
Search is changing in two places at once. On your own site, customers type in their own language and a keyword index hands them nothing. In front of your site, an answer engine now reads the page and states a conclusion, and the click that used to follow the ranking sometimes never happens. Both are the same problem: whether a machine can understand what you sell.
Quick answer
AI search covers two related shifts. On-site, it means semantic and retrieval-based search that ranks products by what a query means rather than by which words it shares with a product title. Off-site, it means Generative Engine Optimization: structuring content so AI answer engines can read it, understand what a business does, and cite it correctly. It matters because both change the same assumption, that a customer types keywords and picks from a ranked list. The business problem AI search addresses is being found and understood by systems that answer rather than list.
Best suited for
- Catalogues large enough that customers search rather than browse
- Businesses whose buyers now start in an AI assistant
- Content-led brands that earn traffic through explanation
- Any site where the no-result page is a real share of sessions
Business problems
The symptom as the person with the problem describes it, what it is costing, and where the technology actually helps.
Customers search and get nothing back
A no-result page is a customer who arrived ready to buy and left. On most catalogues it is a meaningful share of search sessions, and it is the least monitored screen on the site because nobody reports on it.
OpportunityRetrieval-based search that matches meaning, with the keyword index kept underneath for SKUs and exact model numbers. The measurable version is the no-result rate before and after.
AI answers describe the business incorrectly
Language models increasingly sit between a buyer and a brand. Whether they can state what you sell, to whom, and on what terms is now a content structure and structured data problem rather than a public relations one.
OpportunityClear definitions, direct answers, schema that matches the visible page, and an author with a verifiable record. AI systems reward pages that answer a question in one passage, which is also what makes a page useful to a person.
Rankings hold but traffic falls
The position is the same and the click is not. When the answer is stated above the results, the informational half of a content strategy stops converting to sessions even though nothing on the site got worse.
OpportunityShifting the content that exists to earn the citation rather than the click, and moving conversion weight onto the queries where a person still needs a page. This is a measurement change before it is a content change.
Search relevance is a synonym list nobody maintains
Every new product category means another set of hand-written synonyms, and the list rots. The team ends up maintaining a dictionary instead of a catalogue.
OpportunityEmbeddings replace most of that maintenance, because meaning does not need to be enumerated. The synonym list shrinks to genuine business vocabulary rather than every phrasing a customer might use.
Approach
Five stages, in this order, on every engagement. Skipping the first one is the most expensive habit in this industry.
Understand
Business modelCustomersProductsOperations
Read the search logs first. They are the most honest document any business owns: what customers actually call your products, what they expect you to stock, and where they gave up. This costs nothing and usually changes the brief.
Architect
SystemsIntegrationsCustomer journey
Decide what is retrieved and from where. Which fields carry meaning, which need to stay exact, where the vector index lives, and how it stays in step with a catalogue that changes daily. An index that drifts from the catalogue is worse than the keyword search it replaced.
Build
TechnologyExperienceFunctionality
Embeddings over the catalogue, hybrid ranking that keeps exact matching for SKUs and model numbers, and the answer surfaces themselves. Off-site, the content structure and schema that let an answer engine read a page correctly.
Optimize
PerformanceCROSEOCustomer experience
Review the top failing queries every month, because ranking that tests well and behaves badly on a live catalogue is normal. On the GEO side, check what the answer engines actually say about the business, and treat a wrong answer as a content defect with an owner.
Scale
AutomationAIAnalyticsPersonalization
Catalogue enrichment feeding the index, so search improves because the product data improved. Then ranking informed by real behaviour. In that order, because personalising a bad index personalises the wrong results.
Capabilities
On-site search
- Semantic and vector retrieval
- Hybrid keyword ranking
- Faceting and filters
- No-result recovery
- Search analytics
Discovery
- Query intent classification
- Category and collection mapping
- Recommendations
- Merchandising rules
GEO and AI visibility
- Content structure for extraction
- Schema.org and entity clarity
- llms.txt
- Citation monitoring
- Author and record signals
Technical SEO
- Crawlability and indexation
- Canonical and URL structure
- Core Web Vitals
- Structured data validation
Measurement
- No-result rate
- Search-session conversion
- Query coverage
- AI answer accuracy
AI and technology
What each application does, how it works, where the value is, where a person still reviews it, and what it will not do.
Semantic retrieval
- What it does
- Search that ranks by what a query means rather than by which words it shares with the product record.
- How it works
- Product and content text is encoded as vector embeddings; a query is encoded the same way and matched by distance. The keyword index stays underneath and wins on exact identifiers, so a SKU search still behaves like a SKU search.
- Where the value is
- No-result rate and conversion on sessions that used search. Both exist before the project starts, which is what makes this worth funding rather than piloting.
- Where a human reviews
- A merchandiser reads the failing queries monthly. Relevance is a business opinion at the edges, and the model does not hold one.
- What it will not do
- It cannot rescue thin product data. Embeddings of an empty description are an accurate representation of nothing.
Query understanding
- What it does
- Classifying what a customer is trying to do: find a specific item, compare options, or check whether you stock a category at all.
- How it works
- Intent classification over the query stream, mapped onto different result layouts. A comparison query and a known-item query should not return the same page.
- Where the value is
- Fewer sessions that end on a result page nobody engaged with, and a clearer picture of which categories customers expect and you do not carry.
- Where a human reviews
- The category gaps this exposes are a buying decision, not a search decision. It surfaces them and stops.
- What it will not do
- Classification is confident on common queries and unreliable on the long tail, which is exactly where the interesting queries are.
Generative Engine Optimization
- What it does
- Making content that an AI answer engine can read, understand and cite without misrepresenting the business.
- How it works
- Direct answers near the top of a section, definitions that stand alone when extracted, schema that matches the visible text, a named author with a verifiable record, and llms.txt describing what the site is.
- Where the value is
- Being cited correctly when the answer arrives before the link. On this site that is why every FAQ is open text rather than an accordion, and why the structured data is built from the same array the page renders.
- Where a human reviews
- Someone has to read what the answer engines currently say about the business. There is no dashboard that will do this honestly yet.
- What it will not do
- None of it is a ranking control. It makes a page easier to understand and quote, and it cannot make a system cite you. Anyone selling guaranteed AI visibility is selling something they do not have.
Catalogue enrichment for search
- What it does
- Filling in the attributes and descriptions that both the index and the customer need.
- How it works
- A pipeline drafts missing fields against a fixed schema from existing data, feeds and images, and writes back in reviewed batches.
- Where the value is
- Coverage, and search quality as a downstream effect. This is usually the cheapest way to improve search results without touching the search engine at all.
- Where a human reviews
- Review before publish. A fluent, confident, wrong specification on a product page is a returns problem.
- What it will not do
- It works from what it is given. Wrong source data becomes enriched wrong data, at scale and faster.
Architecture
How the pieces sit together, top to bottom. Only technologies actually used appear here.
Customer
Typing in their own language, on your site or in an assistant that answers before it links.
On-site searchAI assistants
Query layer
Understanding the query before deciding what to retrieve.
Intent classificationQuery parsing
Retrieval
Meaning and exactness, blended rather than chosen between.
Vector indexKeyword indexHybrid ranking
Content and catalogue
The source of truth. Everything above is a projection of this, and inherits its gaps.
Product dataAttributesEditorial content
Structure
What makes the same content legible to a machine that is answering rather than listing.
Schema.orgllms.txtSemantic HTML
Measurement
The numbers that decide whether any of it worked.
No-result rateSearch conversionCitation checks
When to use it
Good fit
- A catalogue large enough that customers search rather than browse
- Search logs full of queries that return nothing
- Content that earns its traffic by explaining something
- Product data good enough to be worth retrieving
- Someone in the business who will read the failing queries
Think twice
- A catalogue of a few dozen products, where good navigation beats any search engine
- Product data so thin that retrieval would surface the gaps rather than fix them
- An expectation that GEO controls what an AI system says about you
- No baseline measurement, which makes the result unarguable in both directions
- A business hoping this replaces the technical SEO it has not done
Comparison
These are layers, not rivals. Nearly every working implementation is hybrid: embeddings for meaning, the keyword index for identifiers, and technical SEO underneath both because a page an engine cannot crawl is a page no model will ever cite. Replacing keyword search outright is the most common way this goes wrong.
| Criterion | AI search | Keyword search | Classic SEO |
|---|---|---|---|
| Matches on | Meaning | Exact terms | Terms and links |
| Handles unseen phrasing | Yes | Only with synonyms | Partly |
| Exact SKU lookup | Needs a keyword fallback | Its strongest case | Not applicable |
| Maintenance | Index freshness | Synonym lists | Ongoing content work |
| Setup cost | Moderate | Low | Low to moderate |
| Where it acts | On your site | On your site | In front of your site |
| Measured by | No-result rate, search conversion | Same, on a lower ceiling | Rankings, sessions, citations |
| Fails by | Confident wrong ranking | Returning nothing | Slow, visible decline |
Relevant work
Search and discovery work from the record, and it is technical SEO and catalogue work rather than AI search. That distinction is deliberate: there is no delivered AI search engagement with a published measurement yet, and this section will not borrow one.
Web platform · ManufacturingEngineering manufacturer platform
A modern engineering brand platform built for discovery and enquiry.
Web platformTechnical SEOContent
3x traffic growth and 45% better engagement
Custom platform · B2BProcurement and approvals platform
Scalable B2B procurement platform streamlining purchase requests and approvals.
B2B workflowsApprovalsIntegrations
Business outcomes
3x traffic growth and 45% better engagement
Discovery on an engineering brand platform
Search and content work on a manufacturer platform built for discovery and enquiry rather than for transactions.
Where this came fromFrom the engineering manufacturer build on the work record. It came from technical SEO, site structure and content, before any AI retrieval was involved, and it is quoted here as what search work produced rather than as an AI result.
No-result rate
The most improvable number in on-site search and the least reported. Every no-result session is a customer who told you what they wanted in their own words and got nothing back.
Search-session conversion
Customers who use search convert at a different rate to customers who browse. Separating the two is usually the moment a business starts taking its search seriously.
Answer accuracy
Whether an AI assistant can state correctly what the business sells and to whom. It is checkable by hand today, and treating a wrong answer as a content defect with an owner is most of GEO in practice.
One measured figure here, and it belongs to technical SEO and content work rather than to AI retrieval. The rest are the numbers this work is measured on, listed as numbers to establish rather than results already achieved. There is no AI search engagement on the record with a published outcome, and there will not be a claim of one on this page until there is.
FAQ
- What is AI search?
- Search that ranks results by the meaning of a query rather than by word overlap with the indexed text. It is built on vector embeddings and retrieval, usually blended with a keyword index so exact identifiers still behave correctly. In practice it means a customer can describe what they want in their own words and still find it.
- What is Generative Engine Optimization?
- GEO is the practice of structuring content so AI answer engines can read it, understand it and cite it correctly. It overlaps heavily with good technical SEO and clear writing: direct answers, definitions that survive being extracted from the page, schema that matches the visible text, and a named author with a record behind them.
- Is GEO different from SEO?
- It is a shift of emphasis rather than a separate discipline. SEO optimises to be ranked in a list; GEO optimises to be understood and quoted in an answer. The technical foundation is the same, and a site that cannot be crawled or whose structured data contradicts its page will not be cited by anything.
- Can you guarantee visibility in AI Overviews or ChatGPT?
- No, and neither can anyone else. There is no submission, no ranking control and no published mechanism. What can be done is to make the site easy to read, accurate in its structured data, and clearly attributable to a real person with a record. Anyone selling a guarantee is selling something they do not have.
- How does semantic search actually work?
- Text is converted into vectors that place similar meanings near each other. A query is converted the same way, and the nearest items are retrieved. Because the comparison is on meaning rather than characters, a query the catalogue has never seen phrased that way still matches, which is the part a synonym list cannot do.
- Will AI search replace keyword search?
- No, and implementations that try tend to regress. Keyword matching is better at exact identifiers, model numbers and brand names, where a customer is being precise on purpose. Nearly every working system is hybrid, with meaning and exactness blended and the balance tuned against real queries.
- What does AI search cost to run?
- Embedding a catalogue is inexpensive and one-off; keeping the index in step with a catalogue that changes daily is the ongoing cost, along with the review time to check what is failing. Budgeting for the build and not for the maintenance is the usual mistake.
- How do I measure whether AI search worked?
- Three numbers, all of which exist before you start: the share of searches that return nothing, the conversion rate of sessions that used search, and the share of the query stream that returns something a person then engaged with. If those are not being recorded today, record them for a month before changing anything.
- Does my site need llms.txt?
- It is cheap and it is honest, so this site publishes one. It is a plain description of what the site is and what it covers, at /llms.txt. It is not yet a standard anyone is obliged to read, and it should be treated as a courtesy rather than as an optimisation with a return attached.
- Do you have an AI search case study?
- Not with a published measurement, and I would rather say that here than dress up adjacent work as one. The search and content record is real, including 3x traffic growth on an engineering manufacturer platform, and it came from technical SEO rather than from AI retrieval. The AI layer is where this practice is going, not where its published results already are.
Next step
Let us look at your business model, the systems you already run and where you are trying to get to, before choosing the technology. That conversation is usually shorter than people expect, and it occasionally ends with me saying you do not need the project.